ON/OFF transient ganglion cells of the turtle retina show distinct spike event patterns in response to abrupt intensity changes, e.g. during saccadic eye movements. These patterns consist of two main spike events, with the latency of each event showing a systematic dependence on stimulus contrast. While the latency of the first event decreases monotonically with increasing contrast as expected, the second event shows the shortest latency for intermediate contrasts and a longer latency for high and for low contrasts. These spike event patterns improve the discrimination of different light intensity transitions based on ensemble responses of the ON/OFF transient ganglion cell subpopulation. While the discrimination results are far better than chance using either spike counts or latencies of the first spikes, they are further improved by using properties of the second spike event. The best classification results are obtained when spike rates and latencies of both events are considered in combination. Hence, spike counts and temporal structure of retinal ganglion cells carry complementary information about the stimulus condition, and thus spike event patterns could be an important aspect of retinal coding. To investigate the origin of the spike event patterns in retinal ganglion cells, two computational models of retinal processing are compared. A linear-nonlinear model consisting of separate filters for ON and OFF response components fails to reproduce the spike event patterns. A more complex cascade filter model, however, accurately predicts the timing of the spike events by using a combination of gain control loop and spike rate adaptation.